bghira / terminus-xl-velocity-training

terminus-xl-velocity-training

This is a full rank finetune derived from ptx0/terminus-xl-velocity-v2 .

The main validation prompt used during training was:

a cute anime character named toast holding a sign that says SOON, sitting next to a red square on her left side, and a transparent sphere on her right side

Validation settings

• CFG: 7.5
• CFG Rescale: 0.7
• Steps: 30
• Sampler: euler
• Seed: 42
• Resolutions: 1024x1024,1152x960,896x1152

Note: The validation settings are not necessarily the same as the training settings .

You can find some example images in the following gallery:

The text encoder was not trained. You may reuse the base model text encoder for inference.

Training settings

• Training epochs: 13
• Training steps: 23000
• Learning rate: 4e-07
• Effective batch size: 512
  • Micro-batch size: 32
  • Gradient accumulation steps: 2
  • Number of GPUs: 8
• Prediction type: v_prediction
• Rescaled betas zero SNR: True
• Optimizer: AdamW, stochastic bf16
• Precision: Pure BF16
• Xformers: Enabled

Datasets

photo-concept-bucket

• Repeats: 0
• Total number of images: ~557568
• Total number of aspect buckets: 5
• Resolution: 1.0 megapixels
• Cropped: True
• Crop style: random
• Crop aspect: random

Inference

import torch
from diffusers import DiffusionPipeline



model_id = "terminus-xl-velocity-training"
prompt = "a cute anime character named toast holding a sign that says SOON, sitting next to a red square on her left side, and a transparent sphere on her right side"
negative_prompt = "malformed, disgusting, overexposed, washed-out"

pipeline = DiffusionPipeline.from_pretrained(model_id)
pipeline.to('cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu')
image = pipeline(
    prompt=prompt,
    negative_prompt='',
    num_inference_steps=30,
    generator=torch.Generator(device='cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu').manual_seed(1641421826),
    width=1152,
    height=768,
    guidance_scale=7.5,
    guidance_rescale=0.7,
).images[0]
image.save("output.png", format="PNG")